PROJECT SUMMARY Magnetic resonance imaging (MRI) is a versatile imaging modality that suffers from slow acquisition times which is a challenge for both time sensitive applications and for patient throughput. Accelerating MRI would benefit patients both by reducing the time they need to be in the scanner and in reducing the cost of healthcare. This project is part of a larger scientific effort to accelerate MRI while maintaining the diagnostic quality. Acceleration, even by a factor of two would result in a major advance for public health. Two of the current approaches to accelerate MRI rely on collecting less data (under-sampling) and constrained or deep learning reconstruction. These approaches can lead to images with diagnostic quality with significant under-sampling but may suffer from artifacts which are hard to characterize. Specifically, this project will optimize the performance of constrained reconstruction and deep learning on detecting subtle lesions in acquiring and reconstructing under-sampled MRI. To carry out this optimization, we will first develop the methods required for detection of lesions by machine and human observer models. Then the models will be validated by psychophysical studies where humans perform the detection task. In the first aim of this project, we will optimize constrained reconstruction based on the ideal linear observer. We will consider under-sampled acquisition strategies in 2D MRI including one and two dimensional subsampling methods with constrained reconstruction using both wavelet and total variation constraints. We will perform simulations using anatomical backgrounds both for lesions which match the prior information of the constraints and those which do not to better understand how choices in acquisition and reconstruction affect ideal detection. While the ideal linear observer approximates the best possible detection, typically the signal detection is carried out by a human. In the second aim, we will optimize constrained reconstruction using human observer models and validate the models using human observer studies. A recent approach to reconstruction of under- sampled images is based on deep learning. In the third aim, this work will optimize deep learning reconstruction based on ideal and human observers. Due to the complexity of the deep learning approach, having this task-based approach to optimization is particularly relevant. This project will optimize a network using signal detection to better understand how training and architecture choices in the neural network affect detection of lesions which are not included in training images. This research project will help to strengthen the research environment at Manhattan College by involving students in biomedical research incorporating applied mathematics, statistics and data science.